A Survey on Job and Task Scheduling in Big Data

Transcription

1 A Survey on Job and Task Scheduling in Big Data A.Malcom Marshall 1,Dr.S.Gunasekaran 2 1 PG Scholar (M.E), 2 Prof & Head, 1,2 Dept. of Computer Science & Engineering Coimbatore Institute of Engineering and Technology Coimbatore, India 1 malcom.research@gmail.com 2 gunaphd@yahoo.com Abstract Bigdata handles the datasets which exceeds the ability of commonly used software tools for storing, sharing and processing the data. Classification of workload is a major issue to the Big Data community namely job type evolution and job size evolution. On the basis of job type, job size and disk performance, clusters are been formed with data node, name node and secondary name node. To classify the workload and to perform the job scheduling, mapreduce algorithm is going to be applied. Based on the performance of individual machine, workload has been allocated. Mapreduce has two phases for processing the data: map and reduce phases. In map phase, the input dataset taken is splitted into keyvalue pairs and an intermediate output is obtained and in reduce phase that key value pair undergoes shuffle and sort operation. Intermediate files are created from map tasks are written to local disk and output files are written to distributed file system of Hadoop. Scheduling of different jobs to different disks are identified after completing mapreduce tasks. Johnson algorithm is used to schedule the jobs and used to find out the optimal solution of different jobs. It schedules the jobs into different pools and performs the scheduling. The main task to be carried out is to minimize the computation time for entire jobs and analyze the performance using response time factors in hadoop distributed file system. Based on the dataset size and number of nodes which is formed in hadoop cluster, the performance of individual jobs are identified Keywords hadoop; mapreduce; johnson algorithm; I. INTRODUCTION Bigdata is defined as the large collection of datasets which is complex to process. The organization face difficulties to create, manipulate and manage the large datasets. For example, if we take the social media Facebook, there will be some posts on the page. The number of likes, shares and comments are given at a second for a particular post it leads to creation of large datasets which gives trouble to store the data and process the data. It involves massive volume of both structured and unstructured data. Facebook handles 50 billion photos from its user base. The challenges in bigdata includes include capture, storage, search, sharing, transfer, analysis, and visualization. Bigdata has massive volume of both structured information of data as well as unstructured. Structured data has well defined format where unstructured data does not have well define format. Data mining and Data warehousing technique involves 95% time for gathering data an d only 5% for analyzing the data and no parallel process been carried out. But Bigdata spends 70% time for gathering data and 30% time for analyzing the data.it speed up the complex queries and parallel process are carried out in bigdata platform. Very large datasets are quickly processed by bigdata. Eventhough the data size of data exceeds in petabytes or Exabyte s in range it process the queries with correct accuracy. A. Characteristics II. BIGDATA Bigdata involves three main characteristics namely, Volume Variety Velocity Volume-The quantity of data which is generated is considered as volume which determines the size only. The data size may be in petabytes (1000 terabytes). Variety- The category to which the data belongs is defined by variety. This helps the people who are closely analyzing the data. Heterogeneous, complex, and variable data, which are generated in formats as different as e- mail, social media, video, images, blogs and sensor data. Velocity-It specifies the speed of generation of data or how fast the data is generated. A small Hadoop cluster includes a single master and multiple worker nodes. The master node consists of a Job tracker, Task tracker, Name node and Data node maintaining the Integrity of the Specifications B. Job Tracker (JT) The Job tracker is a node which controls the job execution process, it coordinates all the jobs run on a system by scheduling tasks to run on task trackers. JT performs mapreduce tasks to specific nodes in the cluster. Clients submit jobs to the Job tracker. It informs to the name node to determine the location of the data and locates Task tracker nodes with available slots at or near the data. Jobtracker submits the work to the chosen Task tracker nodes. The Task tracker nodes are monitored and if they do not submit heartbeat signals often, they are failed and the work is scheduled on a different Task

2 tracker. A Task tracker will notify the Job tracker when a task fails. The Job tracker may resubmit the job elsewhere, it may mark that specific record as something to avoid and it may even blacklist the Task tracker as unreliable. When the work is completed, the Job tracker updates its status. Client applications can ask the Jobtracker for information. The Job tracker is a point of failure for the hadoop mapreduce service. If it goes down, all running jobs are halted. C. Task Tracker (TT) A Task tracker is a node in the cluster that accepts tasks. TT run tasks and sends the progress reports to JT, which keeps a record of overall progress of each job. Every Task tracker is configured with a set of slots, it indicate the number of tasks that it can accept. When the job tracker tries to find somewhere to schedule a task within the mapreduce operations, it first looks for an empty slot on the same server that hosts the data node containing the data, and if not, it looks for an empty slot on a ma chine in the same rack. The Task tracker make a separate JVM processes to do the actual work, this is to ensure that process failure does not take down in the task tracker. When the process finishes, the tracker notifies to the job tracker. The task trackers also send out heartbeat messages to the Job tracker, usually every few minutes, to reassure the Job tracker that it is still alive. These message also inform the Job tracker of the number of available slots, so the Job tracker can stay up to date with where in the cluster work can be allocated. D. Name Node (NN) The Name node is the central part of an HDFS file system. It keeps the directory of all files in the file system, and tracks the cluster where the file data is kept. It does not store the data of these files itself. Client applications talk to the Name node whenever they wish to locate a file, or when they want to add, copy, move, delete a file. The Name node is a single point of failure for the HDFS Cluster. HDFS is not currently a high availability system. When the name node goes down, the file system goes offline. There is an optional secondary name node that can be hosted on a separate machine. It only creates checkpoints of the namespace by merging the edits file into the fsimage file and does not provide any real redundancy. E. Data Node (DN) A Data node stores the data in Hadoop file system. A functional file system has more than one data node, with data replicated across them. On startup, a data node connects to the name node; spinning until that service comes up. It then responds to requests from the name node for file system operations. Client applications can talk directly to a data node, once the name node has provided the location of the data. All data nodes send a heartbeat message to the name node every 3 seconds to say that they are alive. If the name node does not receive a heartbeat from a particular data node for 10 minutes, then it considers that data node to be dead or out of service. F. Hadoop distributed file system HDFS has master/slave architecture. An HDFS cluster consists of a single name node, a master server that manages the file system namespace and regulates access to files by clients. In addition, there are a number of data nodes, usually one per node in the cluster, which manages storage attached to the nodes that they run on. HDFS exposes a file system namespace and allows user data to be stored in files. Internally, a file is split into one or more blocks and these blocks are stored in a set of data nodes. The name node executes file system namespace operations like opening, closing, and renaming files and directories. It also determines the mapping of blocks to data nodes. The data nodes are responsible for reserving read and write requests from the file system s clients. The data nodes also perform block creation, deletion, and replication upon instruction from the name node. III. LITERATURE SURVEY In this papers various scheduling of task and the job are surveyed and its descriptions are mentioned: A. Dynamic slot allocation technique [3] 1. Dynamic Hadoop Fair Scheduler Dynamic Hadoop Fair Scheduler (DHFS) schedules the job in order manner; it has pool - independent DHFS (PI- DHFS) and pool - dependent DHFS (PD- DHFS). Pool dependent DHFS have each pool which always satisfy its own map and reduce tasks with its shared map and reduce slots between its map- phased pool and reduce- phased pool,it is known to be intra pool dynamic slots allocation. Pool independent DHFS considers the dynamic slots allocation from the cluster level, instead of pool - level. The map tasks have priority in the use of map slots and reduce tasks have priority to reduce slots are called as intra phase dynamic slots allocation.[3 ]When the respective phase slots requirements met excess slots be used by other phase are referred as inter phase dynamic slots allocation fine. 2. Dynamic slot allocation It is a technique for managing global session resources and allows dynamic resizing of the cache per-client, per-load basis. The client communicates to the server about whether resources are filled in all slots or not filled in the slots. The server then decides how many slots it should allocate to that client in the future. Communication occurs via the sequence operation, which means that updates occur on every step. Suppose if the allocations of resources

3 are free, slots are allocated by using job scheduling. 3. Cluster analysis Clustering refers to grouping of similar objects or data together. Hadoop clusters nodes from a few nodes to extremely large clusters with thousands of nodes. It consists of data node, name node and secondary name node for mapreduce process. The analysis of clusters to different datasets takes place by applying mapreduce technique. 4. Process of Hadoop Fair Scheduler The Fair Scheduler was designed with four main goals. It runs small jobs quickly, even if they are sharing a cluster with large jobs. Provide guaranteed service levels to production jobs, to run jobs in a shared cluster. The users should only need to configure it. Support reconfiguration at runtime, without requiring a cluster restart. Fair scheduling is a method of assigning resources to jobs. When there is a single job running, that job uses the entire cluster. When other jobs are submitted, tasks slots that free up are assigned to the new jobs, so that each job gets roughly the same amount of CPU time. Unlike the default Hadoop scheduler, which forms a queue of jobs, short jobs finish in reasonable time while not starving long jobs. It is also an easy way to share a cluster between multiple of users. Fair sharing can also work with job priorities, the priorities are used as weights to determine the fraction of total compute time that each job gets. The Fair Scheduler can limit the number of concurrent running jobs per user and per pool. This can be useful when a user must submit hundreds of jobs at once, or for ensuring that intermediate data does not fill up disk space on a cluster when too many concurrent jobs are running. 5. Identification of the problem When the Hadoop Fair Scheduler is used in scheduling process, the slot utilization become very low. To address this problem, the slots can be allowed dynamically to either map or reduce tasks depending on their actual requirement. B. Job scheduling for multi-user mapreduce clusters[6] 1. Job scheduling process Job scheduling in Hadoop is performed by the job master, which manages a number of worker nodes in the cluster. Each worker has a fixed number of map and reduce slots, which can run tasks. The workers periodically send heartbeats to the master to reporting the number of free slots and tasks. A job scheduler is a computer application for controlling background program execution. The objective of scheduling is to determine the job schedules that minimize or maximize a measure of performance. It can be characterized by a set of jobs, each with one or more operations. The operations of a job are to be performed in a specified sequence on specific machines. A job scheduler is a program that enables to schedule and monitor computer batch jobs. A job scheduler can initiate and manage jobs automatically by processing prepared job control language statements or through equivalent interaction with a human operator. Job schedulers provide a graphical user interface and a single point of control for all the work in a distributed network of computers. 1. Multi user cluster scheduling The jobs are composed of small independent tasks, it is possible to isolate while using cluster efficiently. The steps in multi user cluster scheduling are to make tasks small in length having small tasks make other new jobs to startup quickly. Delimit the Tasks into pieces with resource requirements. 2. Delay scheduling The job which is to be scheduled next should wait for the previous job to be completed. After completion of the previous job, it should be processed. There occurs a delay in this scheduling of jobs are known as delay scheduling. 3. Copy compute splitting HDFS copy tasks that are fairly CPUintensive because they perform large amounts of memory copies when merging map outputs, so there is little gain from copy compute splitting unless the network bandwidth is very limited (copy tasks compete with compute tasks with CPU). 4. Identification of the problem Sharing a mapreduce cluster between users is attractive because it enables statistical multiplexing (lowering costs) and allows users to share a common large data set. The traditional scheduling algorithms can perform very poorly in mapreduce due to two aspects of the mapreduce setting: the need for data locality and the dependence between map and reduce tasks. C. Automatic resource allocation for mapreduce environment 1. Automatic resource inference It comprises of three inter- related components. First for a production job that is routinely executed on a new dataset. Builds a job profile during mapreduce stages. Third we design a mapreduce performance model and need to estimates the amount of resources required for job completion within the deadline. One of the main features of programs is the amount of resources which are needed in order to run

4 them. Different resources can be taken into consideration, such as the number of execution steps, amount of memory allocated, number of calls to certain methods. Unfortunately, manually determining the resource consumption of programs is difficult and error-prone. 2. SLO scheduler Job scheduling in Hadoop is performed by job master, which manages a number of worker nodes in the cluster. Each worker has a fixed number of map and reduce tasks slots, which can run tasks. The workers periodically sends heartbeats to the master to reporting the number of free slots and schedule the policy. It is based on the observation that we can profile a job that runs routinely and then use its profile in the designed Mapreduce performance model to estimate the amount of resources required for meeting the deadline. These resource requirements are enforced by the SLOscheduler. 3 Resource allocation Resource allocation is the distribution of resources. It is used to assign the available resources. It is part of resource management. In project management, resource allocation is the scheduling of activities and the resources required by those activities while taking into consideration both the resource availability and the project time. Resource allocation may be decided by using computer programs applied to a specific domain to automatically and dynamically distribute resources to applicants. 4 Identification of the problem The key challenge in automatic resource inference is to control resource allocations to different applications for achieving performance goals. Reduce job execution time depends on the amount of map and reduce slots allocated to the job over time. One can speed- up the job completion by increasing the resource allocation and it can be done during both map and reduce stages. This creates new opportunities for different scheduling policies and their performance comparison. D. Interfe rence and Locality- Aware Task Scheduling for Mapreduce Applications in Virtual Clusters 1 ILA Scheduling ILA scheduler performs the interference and locality- aware scheduling operations on top of the fair scheduling. At each interval, it selects a job from a wait queue sorted according to the job s fairness. ILA scheduler always considers the interference mitigation due to the following two reasons:1) VM interference causes much more performance degradation than remote data access. Scheduling a nonlocal task only the individual task. I n contrast, VM interference may affect not only the scheduled task but also all the tasks running on the same VM or physical server. 2) No interference is a precondition for achieving good data locality. Any unexpected task slowdown would make the data locality policy ineffective. 2 Locality aware task scheduling (LARTS) The scheduling algorithm of a runtime determines when and where a computation or a task, executes. Therefore by making the scheduling algorithm locality aware, locality can be exploited. LARTS is a practical method for improving Mapreduce performance. It attempts to collocate reduce tasks with the maximum required data computed after recognizing input data network locations and sizes. It provides a good data locality while there is a scheduling delay, poor system utilization, and low degree of parallelism. 3 Setting up of virtual clusters Hadoop Clusters can be created on Virtual Machines.The process to set up a cluster using the hadoop shell scripts are: All machines should be brought up to date, with the same version of Java, Hadoop. All machine's(both VM's and physical machines) public key are distributed to all "~/.ssh/authorized keys" file.conf/hadoopsite.xml file is similar for all the machines /etc/hosts file must contain all the machines(vm, Physical machine)ip and Hostname. The local hostname entry should be to the assigned IP address.conf/slaves must contain the hostname of all slaves including VM's and physical machine.conf/masters must contain only master's hostname. Both conf/masters and conf/slaves files must be similar in all the participating machines. The hadoop home directory should be same for all the machines. Finally execute the namenode format command only in the Master machine. 4 Identification of the problem Interference aware scheduling policy is based on the task performance prediction model an adaptive scheduling algorithm for data locality improvement. It improves the data locality of map tasks upto 65%.Virtualisation affect the performance of mapreduce applications. In virtual mapreduce cluster, the interference may cause variation in virtual machine capacity. In mapreduce clusters, locally running tasks, HDFS may also issue I/O requests for a remote data access. ILA considers this I/O as the background traffic and manages to mitigate the interference through task scheduling. Extend ILA scheduler to mitigate the

5 interference from HDFS through intelligent data placement and data node selection. E. Improving mapreduce performance in heterogeneous Environments [7] 1 Heterogeneity Hadoop assumes that any detectably slow node is faulty. But nodes can be slow for other reasons. In a non - virtualized data center, there may be multiple generations of hardware. In a virtualized datacenter where multiple virtual machines run on each physical host, such as Amazon EC2, co- location of VMs may cause heterogeneity. Although virtualization isolates CPU and memory performance, VMs compete for disk and network bandwidth. Heterogeneity seriously impacts Hadoop s scheduler. Because the scheduler uses a fixed threshold for selecting tasks to speculate, too many speculative tasks maybe launched, taking away resources from useful tasks. Also, because the scheduler ranks candidates by locality, the wrong tasks may be chosen for speculation first. 2 The LATE Scheduler [Longest Approximate Time to End] The primary task provides the greatest opportunity for a speculative copy to overtake the original and reduce the job s response time. This greedy policy would be optimal if nodes ran at consistent speeds and if there was no cost to launching a speculative task on an otherwise idle node LATE takes into account node heterogeneity when deciding where to run speculative tasks. Hadoop s native scheduler assumes that any node that finishes a task and asks for a new one is likely to be a fast node, that slow nodes will never finish their original tasks and so will never be candidates for running speculative tasks. This is clearly untrue when some n odes are only slightly slower than the mean. Finally, by focusing on estimated time left rather than progress rate, LATE speculatively executes only tasks that will improve job response time, rather than any slow tasks. The progress rate based scheduler would always re- execute tasks from slow nodes, wasting time spent by the backup task if the original finishes faster. 3 Speculative Execution in Hadoop When a node has an empty task slot, Hadoop chooses a task for it from one of three categories. First, any failed tasks are given highest priority. This is done to detect when a task fails repeatedly due to a bug and stop the job. Second, non- running tasks are considered. For maps, tasks with data local to the node are chosen first. Finally, Hadoop looks for a task to execute speculatively. To select speculative tasks, Hadoop monitors task progress using a progress score between 0 and 1.For a map, the progress score is the fraction of input data read. For a reduce task, the execution is divided into three phases as, The copy phase, when the task fetches map outputs. The sort phase, when map outputs are sorted by key. The reduce phase, when a user- defined function is applied to the list of map outputs with each key. In each phase, the score is the fraction of data processed. 4 Identification of the problem Hadoop is an open- source implementation of mapreduce and is often used for short jobs where low response time is critical. Hadoop s performance is closely tied to its task scheduler, which implicitly assumes that cluster nodes are homogeneous and tasks make progress linearly and uses these assumption is to decide when to speculatively re-execute tasks that appear to be stragglers. Longest Approximate Time to End that is highly robust to heterogeneity and it can improve Hadoop response times by a factor. E. Prefetching and Pre -shuffling in Shared Mapreduce Computation Environment [8] 1 Prefetching Scheme HPMR provides a prefetching scheme to improve the degree of data locality. The prefetching scheme can be classified into two types: the intra- block prefetching and the interblock prefetching. The intra- block prefetching is a simple prefetching technique that prefetches data within a single block while performing a complex computation. While a complex job is performed the required data are prefetches and assigned in parallel to the corresponding task. The intra- block prefetching has a couple of issues that need to be addressed. First, it is necessary to synchronize computation with prefetching. We have solved this problem using the concept of processing bar that monitors the current status of each side and invokes a signal if synchronization is about to be broken. The processing bar not only works as a synchronization manager, but also provides a way to measure the effectiveness of this technique. The second issue is to find the proper prefetching rate. The inter- block prefetching runs in block level, by prefetching the expected block replica to a local rack. If the local node does not have enough space, it is replicated to one of nodes in the local rack. When replicating

6 a data block by the inter - block prefetching, it is read from the least loaded replica so as not to minimize the impact on the overall performance. 2 Optimized Scheduler The optimized scheduler is a flexible task scheduler with the predictor and the D- LATE module. In order to understand the optimized scheduler, one must first understand the native Hadoop scheduler. The native Hadoop scheduler assigns tasks to available task trackers by checking heartbeat messages from task trackers. Every task tracker periodically sends a heartbeat message to the job tracker. The job tracker has a task scheduler which actually schedules tasks. First, it constructs a list known as Task In Progress List, a collection of all running tasks, and caches the list into either on Running Map (for map tasks) or non-running Reduce (for reduce tasks) depending on the type of each task. These cache disks are used for the job tracker to manage current map tasks or reduce tasks to be executed. Next, task trackers periodically send a heartbeat message to the job tracker using heartbeat message protocol. Finally, the scheduler assigns each task to a node randomly via the same heartbeat message protocol. The best case of scheduling a task is when the scheduler locates the corresponding task into the local node. Hadoop assumes that all cluster nodes are dedicated to a single user, failing to guarantee high performance in the shared Mapreduce computation environment. The prefetching scheme exploits data locality, while the preshuffling scheme significantly reduces the network overhead required to shuffle keyvalue pairs. [7] Joseph.A, Konwinski.A, Katz.R, Stoica.I and Zaharia.M Improving mapreduce performance in heterogeneous environments, in the proceedings of the 8th symposium on operating systems Design and Implementation DEC, [8] Sangwon Seo and Seungryoul Maeng, HPMR Prefetching in Shared mapreduce computation environment, IEEE REFERENCES [1] Abhishek Verma, Ludmila Cherkasova and Roy H. Campbell, Fellow, Orchestrating an Ensemble of Mapreduce Jobs for minimizing their makespan IEEE transactions on dependable and secure computing, vol. 10, no. 5, september/october 2013 [2] Bikash Sharma, Chita R.Das, Mahmut T.Kandemir and Seung- Hwan Lim, MROrchestrator:A Fine- Grained Resource Orchestration for Hadoop mapreduce Jan 2012 [3] Bu- Sung Lee, Bingsheng He and Shanjiang Tang Dynamic slot allocation technique for mapreduce clusters IEEE [4] Campbell.R,Cherkasova.L and Verma.A ARIA:Automatic Resource Inference and Allocation for Mapreduce e nvironments, Conference on Autonomic Computing (ICAC), [5] Cheng- Zhong Xu, Jia Rao and Xiangping Bu, Interference and Locality- Aware Task Scheduling for Mapreduce Applications in Virtual Clusters, [6] Dhruba Borthakur, Ion Stoica,Joydeep Sen Sarma and Scott Shenker, Job scheduling for multi - user mapreduce clusters April 30, 2009.